STFNet: A spatial-temporal feature aggregation network for breast lesion segmentation in ultrasound videos

Accurate segmentation of lesions in ultrasound is critical for early detection and diagnosis of breast cancer. Previous studies have mainly focused on static 2D images, ignoring the clinical scenarios in which sonographers make a diagnosis based on continuous and dynamic scanning results. Ultrasound video can provide more detailed information about the lesion and its surrounding tissues. However, the automatic segmentation of breast lesions in ultrasound videos presents challenges due to the complex spatial and temporal variations within the images. To address this, we introduce STFNet (Spatial-Temporal Feature Aggregation Network), designed for robust segmentation in breast ultrasound videos. STFNet integrates the spatial feature extraction module (SFEM) for local spatial feature extraction and the temporal feature extraction module (TFEM) for temporal dependency modeling, effectively capturing both fine-grained details and long-range contextual dynamics. A dedicated Multi-Scale Feature Fusion (MSFF) module hierarchically combines multi-resolution spatial and temporal features to enhance lesion boundary delineation, while a hybrid loss function incorporating boundary-aware optimization mitigates the impact of speckle noise and low contrast. We validated STFNet in two academic hospitals with 490 patients (550 videos). As expected, STFNet gets Dice scores of 80.27% (Dataset A) and 78.68% (Dataset B), surpassing existing state-of-the-art methods. Our method combine video analysis with clinical practice, offering accurate and automated breast cancer diagnosis.